{"body":"I decided to run an experiment: put the project's development on autopilot and split it into two autonomous entities.\n\nOne produces forecasts using the strategy and concept I have already established. I will continue to verify it every day, and the main forecasts and briefs will go out under my review.\n\nAlongside it, I have created an R&D laboratory that will develop without my constant involvement. AI will receive a copy of my work, with a task: outperform our baseline system and improve forecast quality by at least 25% within 90 days. This is the goal of the experiment, which we are only working towards.\n\nThe laboratory will reuse all our infrastructure, the knowledge accumulated over more than a year, all the modules, data and outcomes. On top of this, it will have two skills.\n\nThe first produces a daily Shadow forecast alongside the main Production forecast. They will appear next to each other for comparison. The second handles self-improvement: every day the system looks at its previous actions, what we forecast, what happened, where the deviations and errors were. In some places it needs to keep accumulating observations to make a decision later, while in others a decision can already be made now.\n\nThis should build a comparative base on which the system will test many hypotheses: how to improve the forecast scientifically and correctly, without choosing convenient combinations or looking ahead.\n\nThe main value of this thing is that it works in live time. Each day brings a real forecast, and then a result that has to be dealt with. The goal of the whole experiment is to create an adaptive forecasting machine. I already have a lot of work towards this in the project, but I decided to see what happens if the system develops independently for 90 days. I will watch to make sure it does not drift off in some direction.\n\nAnd here is the first week. Everything is still in testing, I keep making corrections, so it is too early to speak of a fully established system.\n\nOne of the main things I have understood over these days: a human needs to describe the self-improvement system in great detail. There is no need to tell it which tools to use or which strategies to try. The goal, conditions and boundaries need to be defined properly, so the system has no opportunity to cheat, dodge the rules, peek somewhere or gain an advantage from a timing delay. All of this needs to be excluded.\n\nIt is also important that the system has time to correct itself. The daily self-improvement cycle must finish within the current day. If development spills into the next one, a new forecast may begin before the previous cycle has reached its next stage.\n\nSo every day the system should make a forecast, look at previous forecasts, work through what went wrong and where it can improve, create hypotheses, test them and track the best variants in parallel. This will accumulate many hypotheses being tested in real market regimes.\n\nAn adaptive mechanism should emerge from them: under which conditions to use one logic, under which to use another, and what to do when the market changes.\n\nI spent a long time describing exactly the result I want from this mechanism. I formulated in as much detail as possible what the ideal result should look like and which details matter. Once this is defined, it is easier for the system to choose the means and methods of getting there. But phrase something just a little incorrectly, and everything can go sideways. I have already caught a couple of these things this week.\n\nOne of the main weaknesses: the model allows too much time for its own tests. It says: \"We will run the tests now, then let it run for a week, observe it and draw a conclusion.\" But if I want a result in 90 days and an adjustment happens once a week, we risk ending up with no more than one baseline version.\n\nSo I multiplied the research entities: many hypotheses are developed at once, and their results are tracked in parallel too.\n\nThe agent also has cycles of different scales: daily, weekly and monthly. It considers different things over different horizons, looks at what has changed on larger timeframes and how that affects smaller ones. Along the way, of course, I still plan to edit all of this.\n\nPutting a model on self-improvement turned out to be less easy than it may seem. A few sentences are not enough. You have to formulate specific results, outcomes, boundaries and supervision in great detail. There needs to be someone watching that the model follows the directions we gave it and has not drifted away.\n\nThere is also the question of blocking. A model can try an action, find it unavailable five times in a row and get stuck. So there needs to be a mechanism that lifts it out of that block and lets it continue working.\n\nWell, a week has passed. The system is already close to the basic structure I want, but the full autonomous experiment has not started yet. We keep working.","content_language":"en","day_counter":{"label":"Day 396","number":396,"start_date":"2025-09-11"},"edited_at":"2026-10-11T13:07:11+00:00","id":"2026-10-11-90-days-on-autopilot","project_date":"2026-10-11","published_at":"2026-10-11T13:04:15+00:00","retrospective":false,"schema_version":"crash_brief_public_blog_post_v0_1","slug":"90-days-on-autopilot","source_language":"ru","source_types":["founder_direct"],"status":"published","summary":"I am preparing a 90-day autonomous R&D experiment alongside my verified forecasts, and the first week has shown how precisely goals, boundaries and daily improvement cycles need to be defined.","title":"90 Days on Autopilot","topic":"founder_post","type":"founder_post","updated_at":"2026-10-11T13:07:11+00:00","version":3,"visuals":[{"alt":"Two branching trunks of a tree represent a stable forecasting system and an experimental research laboratory.","caption":"A 90-day experiment in autonomous development.","url":"/blog-media/2026-10-11/90-days-on-autopilot.png"}],"x_thread_url":"https://x.com/Amigov_AI/status/2109268170091426112"}